Multistream CNN-BiLSTM Framework for Enhanced Human Activity Recognition Leveraging Physiological Signal
Abisek Dahal, Soumen Moulik · IEEE Sensors Letters · 2025
Human Activity Recognition (HAR) and classification is one of the most hyped and trending domains in the last decade. HAR involves multiple hit and trial approaches, machine and deep learning have emerged as excellent techniques for analyzing various physiological sensors used to capture human activities. This work introduce a multi-stream CNN-BiLSTM framework that works on physiological signals corresponding to different activities, in order to achieve an enhanced HAR System. In this work EMG signals that capture the muscles data during activities are used to classify various activities. We achieve an overall average of98.06%accuracy in predicting activities. In addition to that we achieve10-20%more as compared to benchmark model in similar dataset with less computational time. Further the proposed model demonstrates better and remarkable performance in HAR 8-channel benchmark SOTA dataset.